Prior reservation, args.prior_loss_weight doesn't exist about sd_dreambooth_extension

[d8ahazard]

When I use prior reservation, it tells me args.prior_loss_weight doesn't exist in the train_dreambooth.py line 796 loss = loss + args.prior_loss_weight * prior_loss

I quickly turned it into loss = loss + prior_loss just to test it and it seems to be churning well. I'm not sure where args.prior_loss_weight should be defined.

also it would be good to specify the labels correctly. db_train_batch_size = gr.Number(label="Batch Size", precision=1, value=1) db_sample_batch_size = gr.Number(label="Class Batch Size", precision=1, value=1) it's confusing, I needed to look into the code to see what is what. so Batch size should be labeled Train Batch Size etc...

This should have been fixed last night, at least the missing prior loss weight value.

As far as "confusing labels" go, I'll be writing up a readme for this today so it's not so esoteric. The problem with coming up with labels is that even the ones within the project don't necessarily line up with various terminology. For example, in one spot, the images used with Prior Preservation are referred as classifier images, but in other parts of the code, they're called "sample" images.

[MV799]

Thanks!
It successfully finished last night, but with the default and about steps 100x images, it brutally overtrained. I'm not sure where was the issue, I'll have to check. I'll try to do another batch today with the new pull.
Thanks, for making this - in general, it seems to be working well.

[d8ahazard]

Thanks! It successfully finished last night, but with the default and about steps 100x images, it brutally overtrained. I'm not sure where was the issue, I'll have to check. I'll try to do another batch today with the new pull. Thanks, for making this - in general, it seems to be working well.

My pleasure.

Try ensuring you have a class prompt and around 100 for the class images as well, see if that helps with the overtraining.

Closing.